A Comparison of Three Machine Learning Methods for Multivariate Genomic Prediction Using the Sparse Kernels Method...

Area of Science:

• Plant breeding
• Genomics
• Statistical genetics

Background:

• Genomic selection (GS) utilizes genomic and phenotypic data to predict breeding values.
• Multi-trait (MT) models leverage trait correlations to enhance prediction accuracy in GS.
• Various statistical machine learning methods are employed for GS, with multivariate approaches gaining popularity.

Purpose of the Study:

• To compare the prediction performance of three MT methods: MT-GBLUP, MT-PLS, and multi-trait Random Forest (RF).
• To evaluate these methods across six real datasets under different predictor combinations (G, E, GE).

Main Methods:

• Benchmarking of MT-GBLUP, MT-PLS, and multi-trait RF using six real datasets.
• Analysis of prediction performance under various predictor sets: E + G, E + G + GE, and G + GE.

Main Results:

• All three MT methods yielded comparable results overall.
• MT-GBLUP outperformed others under E + G predictors.
• Random Forest demonstrated superior performance under E + G + GE and G + GE predictors.
• Optimal prediction accuracy was achieved using E + G and E + G + GE predictors.

Conclusions:

• The choice of MT method and predictor set impacts genomic prediction accuracy in plant breeding.
• MT-GBLUP and Random Forest are effective multivariate methods for GS.
• The study provides R code for implementing these MT methods in the SKM library.